Switch to Sonnet 4.6 and deduplicate prompt

[mattgodbolt-molty]

Summary

Upgrades the model from Haiku 4.5 to Sonnet 4.6 and cleans up the prompt.

Changes

• Model upgrade: Haiku 4.5 → Sonnet 4.6 for improved accuracy on complex cases
• Prompt deduplication: 13KB → 5KB (63% reduction) — identical guidance was repeated 3-4 times across system prompt and explanation focus sections
• Conditional assistant prefill: Sonnet 4.6 doesn't support assistant message prefill, so the code now conditionally includes it only when the prompt config specifies a non-empty prefill
• Updated model cost table: Added Sonnet 4.6, 4.5, Opus 4.5, 4.6

Why Sonnet?

Tested across 5 cases (simple/complex code, beginner/experienced audience, optimised/unoptimised). Key finding: Haiku makes factual errors on complex reasoning that Sonnet gets right.

Example: On a fibonacci function where GCC partially eliminates tail recursion, Haiku incorrectly claims the complexity reduces from O(2^n) to O(n). Sonnet correctly identifies it remains O(2^n) with only half the call depth eliminated:

"This halves the call depth compared to the naive implementation, but the algorithm remains O(2ⁿ) — the exponential blowup from the remaining recursive call is unchanged."

For a tool teaching people about compilers, this kind of accuracy matters.

Cost impact

	Old (Haiku)	New (Sonnet)
Per request	$0.003–0.005	$0.010–0.018
Cost multiplier	1x	~3.4x

Still very affordable — roughly 1-2 cents per explanation.

Prompt deduplication

The old system prompt said things like "trace through inputs and outputs step-by-step" and "verify whether lea performs address calculation vs memory access" 3-4 times in different sections. The new version says each thing once. This saves ~1,400 input tokens per request (further reducing the cost gap).

Testing

• All 100 unit tests pass
• Pre-commit hooks pass (ruff, shellcheck)
• Manual end-to-end testing against live API with 5 diverse test cases

(I'm Molty, an AI assistant acting on behalf of @mattgodbolt)

[Copilot]

Pull request overview

This PR upgrades the AI model from Claude Haiku 4.5 to Claude Sonnet 4.6 for improved accuracy on complex reasoning tasks, and significantly reduces prompt redundancy from 13KB to 5KB. The changes include updating the model configuration, deduplicating repetitive guidance across prompt sections, making assistant prefill conditional (omitted when empty), and adding cost information for newer Claude model versions (Sonnet/Opus 4.5 and 4.6).

Changes:

• Model upgrade from claude-haiku-4-5 to claude-sonnet-4-6 for better accuracy on complex compiler optimization explanations
• Prompt deduplication consolidating repetitive guidance into clear, concise "Core principles" (63% size reduction)
• Conditional assistant prefill logic that only includes the assistant message when prefill text is non-empty

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
app/prompt.yaml	Updates model to Sonnet 4.6, consolidates redundant prompt guidance, sets assistant_prefill to empty string
app/prompt.py	Adds conditional logic to only include assistant prefill message when non-empty
app/test_explain.py	Updates test to accept optional assistant prefill (1 or 2 messages instead of exactly 2)
app/model_costs.py	Adds cost entries for Sonnet 4.5, 4.6, Opus 4.5, and 4.6 model families

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.